Perceptual Transform Coding
The coding of audio utilizes coding techniques that exploit various perceptual models of human hearing. For example, many weaker tones near strong ones are masked so they do not need to be coded. In traditional perceptual audio coding, this is exploited as adaptive quantization of different frequency data. Perceptually important frequency data are allocated more bits and thus finer quantization and vice versa.
For example, transform coding is conventionally known as an efficient scheme for the compression of audio signals. In transform coding, a block of the input audio samples is transformed (e.g., via the Modified Discrete Cosine Transform or MDCT, which is the most widely used), processed, and quantized. The quantization of the transformed coefficients is performed based on the perceptual importance (e.g. masking effects and frequency sensitivity of human hearing), such as via a scalar quantizer.
When a scalar quantizer is used, the importance is mapped to relative weighting, and the quantizer resolution (step size) for each coefficient is derived from its weight and the global resolution. The global resolution can be determined from target quality, bit rate, etc. For a given step size, each coefficient is quantized into a level which is zero or non-zero integer value.
At lower bitrates, there are typically many more zero level coefficients than non-zero level coefficients. They can be coded with great efficiency using run-length coding. In run-length coding, all zero-level coefficients typically are represented by a value pair consisting of a zero run (i.e., length of a run of consecutive zero-level coefficients), and level of the non-zero coefficient following the zero run. The resulting sequence is R0,L0,R1,L1 . . . , where R is zero run and L is non-zero level.
By exploiting the redundancies between R and L, it is possible to further improve the coding performance. Run-level Huffman coding is a reasonable approach to achieve it, in which R and L are combined into a 2-D array (R,L) and Huffman-coded. Because of memory restrictions, the entries in Huffman tables cannot cover all possible (R,L) combinations, which requires special handling of the outliers. A typical method used for the outliers is to embed an escape code into the Huffman tables, such that the outlier is coded by transmitting the escape code along with the independently quantized R and L.
When transform coding at low bit rates, a large number of the transform coefficients tend to be quantized to zero to achieve a high compression ratio. This could result in there being large missing portions of the spectral data in the compressed bitstream. After decoding and reconstruction of the audio, these missing spectral portions can produce an unnatural and annoying distortion in the audio. Moreover, the distortion in the audio worsens as the missing portions of spectral data become larger. Further, a lack of high frequencies due to quantization makes the decoded audio sound muffled and unpleasant.
Wide-Sense Perceptual Similarity
Perceptual coding also can be taken to a broader sense. For example, some parts of the spectrum can be coded with appropriately shaped noise. When taking this approach, the coded signal may not aim to render an exact or near exact version of the original. Rather the goal is to make it sound similar and pleasant when compared with the original. For example, a wide-sense perceptual similarity technique may code a portion of the spectrum as a scaled version of a code-vector, where the code vector may be chosen from either a fixed predetermined codebook (e.g., a noise codebook), or a codebook taken from a baseband portion of the spectrum (e.g., a baseband codebook).
All these perceptual effects can be used to reduce the bit-rate needed for coding of audio signals. This is because some frequency components do not need to be accurately represented as present in the original signal, but can be either not coded or replaced with something that gives the same perceptual effect as in the original.
In low bit rate coding, a recent trend is to exploit this wide-sense perceptual similarity and use a vector quantization (e.g., as a gain and shape code-vector) to represent the high frequency components with very few bits, e.g. 3 kbps. This can alleviate the distortion and unpleasant muffled effect from missing high frequencies and other large portions of spectral data. The transform coefficients of the “missing spectral portions” are encoded using the vector quantization scheme. It has been shown that this approach enhances the audio quality with a small increase of bit rate.
Nevertheless, due to the bit rate limitation, the quantization is very coarse. While this is efficient and sufficient for the vast majority of the signals, it still causes an unacceptable distortion for high frequency components that are very “tonal.” A typical example can be the very high pitched sound from a string instrument. The vector quantizer may distort the tones into a coarse sounding noise.